Skills

Skills for coding agents

A simple way to package processes, habits, and tiny checklists so your agents can repeat the work better.

What is a skill?

A skill is a reusable instruction set for an agent. Think of it as a small operating guide: when a situation shows up, the agent loads that process and follows the steps.

In practice, it is usually a folder with a SKILL.md file containing a name, a description, and the instructions the agent should follow.

skills/
  quality-gate/
    SKILL.md

What is it for?

Skills work best when they package a small, clear, repeatable behavior.

Quality gates

Run lint, tests, build, or type checks before handing a code change back.

PR review

Review pull requests using the team's criteria, not just loose preferences.

Handoff

Prepare a clear summary of what changed, risks, validation, and next steps.

Start with the minimum: instructions, small examples, and validation criteria. If the skill grows, move details into supporting files inside the skill folder. That lets the agent load the basics first and pull in extra resources only when needed.

How does a skill get activated?

A skill is not a hard guarantee. It is a capability the agent can use. Most of the time, the agent sees the skill name and description, decides whether it matches the task, and only then loads the full SKILL.md.

That makes the description very important. It should say what the skill does and when to use it. If the trigger is vague, the agent may simply continue without using the skill.

Implicit

The agent chooses the skill when the task appears to match its description.

Explicit

You call the skill by name, when the tool supports slash commands or direct invocation.

Reinforced

A project rule reminds the agent to consider the skill at important moments.

Claude Code, Cursor, and the harness

Claude Code treats skills as capabilities the agent can discover. You can let Claude choose automatically, or invoke a skill directly with /skill-name. Claude Code also has frontmatter controls for manual-only or agent-only skills.

Cursor also supports Agent Skills, but its strongest mental model still goes through Rules. Rules are persistent context for the Agent, and they can be always applied, attached by file patterns, requested by the agent, or called manually.

In practice: if you want an implicit skill to fire more consistently in Cursor, pair it with a short project rule. The rule does not need to duplicate the whole skill. It only needs to remind the harness when to consider it.

---
description: Use the quality-gate skill after technical code tasks.
alwaysApply: true
---

After code edits, remind the agent to use the quality-gate skill
before handing the task back.

Sharing skills

There are two useful paths for distributing skills: a simple npx command, or marketplaces/plugins when you need discovery, updates, and scale.

Repo + NPX

For getting started fast

Vercel created the skills package: a CLI that installs, lists, updates, and removes skills for different coding agents. You do not publish your skill to npm. npm gives you the skills tool; your skill still lives on GitHub.

npx runs the "skills" CLI
skills reads the GitHub repo
skills looks for SKILL.md files
skills installs the skill into the agent

The CLI accepts owner/repo, a full GitHub URL, a direct path to a skill, GitLab, any git URL, and even a local folder while you are testing.

Copy and paste

Ready-to-use skills

Starting small: one skill that reminds the agent to validate the project before handing code back.

quality-gate

At the end of a technical task, the agent should find and run the best quality gate available: lint, tests, type check, build, or formatting.

Command ready

npx skills@latest add baltazarparra/ai-native-engineering --skill quality-gate --agent codex --global

References

Community

Ask, answer, get unstuck

Use this space to ask questions about the session, share examples, and help other people understand the topic.